- winsorize at 0.95 shown below (seems better than 0.9 and 0.99, which aren't shown below) - badness thresholds: 40 to 80, steps of 5 - outcomes: imputed AFM, FC, MBFC, PC1 (about 11.5k domains) - model: `y_t1 ~ condition * (t0 + expose)` - `condition`: -0.5, 0.5 - `t0`: pre-treatment baseline (sqrt-transform then mean-centered) - `expose`: bad domain exposure via network/friends (sqrt-transform then mean-centered) - below, I also compared four ways of computing this variable: count/sum, count/sum adjusted for proportion of user's friends we manage to sample - not showing `t0` and `expose` effects below because they're highly significant - fixed effects: block and day (8 days) - cluster SE on block # outcome: count badness (quasipoisson model) ![[_count_winsorize-0.95-clustse 4.png]] # outcome: summed badness (quasipoisson model) ![[_sum_winsorize-0.95-clustse 4.png]] # outcome: fraction badness (OLS model) - could also use quasipoisson model given the skewed distribution (somewhat resembles summed badness distributions) ![[_frac_winsorize-0.95-clustse 4.png]]